Claude
7c3c59f5f8
Target selection (SPEC §6): comrak 0.53.0 (45c1995, BSD-2-Clause) and tokei 14.0.0 (fa44e51, MIT/Apache-2.0); both CPU-bound with pinned corpora (progit book 11MB; 8x replicated pinned trees, 221MB) and committed MANIFEST.sha256 + golden-output pins + regen scripts. Calibration: A/A null verdicts on both real workloads; taskset -c 2 pinning tightens the A/A CI from [0.938, 1.008] to [0.993, 1.008], making the 2% accept threshold meaningful in this container. Profiles (callgrind; perf unavailable): comrak allocator symbols 26.7% of Ir; tokei perform_multi_line_analysis 31.7%. Attempts (ledger + results/phase0/): - comrak class-1 target-cpu=native: null [0.983, 1.011], rejected-bench - comrak class-2 mimalloc swap: +4.6% [1.033, 1.058], 848 upstream tests green, golden byte-identical, ASan-proper clean; LSan flags 484B in onig (reachability false-positive analysis in case study) -> needs-human-review per SPEC §8 - tokei class-1 fat LTO+cgu=1: +1.6% [1.003, 1.026], real but under the 2% bar, rejected-bench Also: ledger-cli for manual rows, targets/ excluded from trust-layer workspace, corpora/build artifacts gitignored, case study written. Phase 0 exit criteria (>=10% verified win) not yet met; next hypotheses logged in results/phase0/attempts.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EcKPWQ3MQD4NdEHwar96u2
2.7 KiB
peltier.io
Spot cooling for hot paths. A profile-guided optimization agent that only ships wins it can prove.
hotpath finds the 10–90% left on the floor between "the compiler" and "hire a perf consultant," with two properties no consultant offers: verified equivalence (a change provably preserves behavior or is discarded) and self-verifying ROI (the value claim is a stopwatch on trusted infrastructure, not a survey).
Full design: SPEC.md.
Layout
crates/ Trust layer (agent has NO write access — SPEC §10)
bench-runner/ Interleaved A/B timing, bootstrap CIs, A/A calibration
diff-test/ Equivalence gates, corpus hash-pinning, per-target policy
ledger/ Append-only SQLite attempt ledger (enforced by triggers)
report/ ROI math: speedup CI → cores → dollars, caveats included
agent/ Untrusted proposer (Claude Agent SDK, Python)
playbook/ Optimization classes 1–7, tried strictly cheapest-first
config/ accept.toml (thresholds), pricing.toml (ROI inputs)
targets/ Per-target checkouts — the only agent-writable path
corpora/ Hash-pinned golden-replay inputs (read-only to agent)
results/ Calibration evidence, per-engagement outputs
Quick start
cargo test --workspace # trust-layer unit tests
just aa # A/A self-test: must yield a null verdict
just compare "cmd-a" "cmd-b" # interleaved A/B with bootstrap CI
Status
Phase 0 dry run in progress (see SPEC §5 for the phase plan and exit
criteria). Two targets vendored and pinned (comrak 0.53.0, tokei
14.0.0); A/A calibration passed on both workloads; three attempts in
the ledger — best so far +4.6% [95% CI +3.3%, +5.8%] on comrak via
allocator swap, needs-human-review (sanitizer flag, analysis in
results/phase0/case-study-comrak.md). Exit criteria (≥10% verified
win) not yet met.
Done:
- Workspace with the four trust-layer crates; core statistics (interleaved scheduling, bootstrap ratio CIs, CI-lower-bound accept rule), append-only ledger schema with mutation-refusing triggers, corpus hash-pin verifier, equivalence-policy parser, ROI formulas — all unit-tested.
- bench-runner CLI: whole-program
compareandaamodes with env fingerprinting. - Playbook classes 1–7; agent tool contract and prompting spine.
Next (Phase 1 exit criteria):
perf statcounters + max RSS + RAPL capture per run.- Automated A/A calibration sessions and regression-injection self-test
recorded to
results/calibration/(<5% FP, ≥95% detection). - diff-test orchestration end-to-end on two real targets; harness IPC for the agent tool surface.
License
GPL-3.0-or-later — see LICENSE.